-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrainer.py
356 lines (288 loc) · 15.3 KB
/
trainer.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
import re
import torch
import logging
import numpy as np
from tqdm import tqdm
from tabulate import tabulate
from utils import model_directory
class TrainerBART:
def __init__(self, tokenizer, model, loss, device, model_path, type_model, max_len):
self.device = device
self.tokenizer = tokenizer
self.model = model
self.type_model = type_model
self.loss = loss
self.model_path = model_path
self.max_len = max_len
self.best_mrr = 0
self.patience = 0
def training(self, optimizer,train_dataloader, dev_dataloader, epochs):
logging.info('Starting training')
steps_done = 0
for epoch in range(epochs):
logging.info(f'Epoch {epoch + 1}/{epochs}')
self.model.train()
train_loss_n = 0.0
train_loss_d = 0
iterator = tqdm(train_dataloader)
for batch in iterator:
source = batch["source"].to(self.device)
target = batch["target"].to(self.device)
decoder_input_ids = self.model.shift_tokens_right(target, self.tokenizer.pad_token_id)
optimizer.zero_grad()
logits=self.model(source, decoder_input_ids = decoder_input_ids)[0]
logits = logits.reshape((logits.shape[0]*logits.shape[1],logits.shape[-1]))
gold = target.view(-1)
loss = self.loss(logits, gold)
loss.backward()
optimizer.step()
train_loss_n += loss.item()
train_loss_d += 1
steps_done += 1
iterator.set_postfix(
loss=(train_loss_n / train_loss_d),
epoch=epoch,
step=steps_done
)
training_metrics = {}
training_metrics['loss'] = train_loss_n / train_loss_d if train_loss_d > 0 else 0.0
validation_metrics, self.best_mrr, self.patience = self.evaluate(dev_dataloader, self.best_mrr, self.patience)
keys = set()
keys |= set(training_metrics.keys())
keys |= set(validation_metrics.keys())
table = []
for key in keys:
table.append((key, training_metrics.get(key, float('Nan')), validation_metrics.get(key, float('Nan'))))
print(tabulate(table, headers=('metric', 'train', 'dev')) + '\n')
if self.patience ==3:
print("\033[1m No improvement for 3 epochs in a row, stop \033[0m \n")
break
logging.info('Complete')
def evaluate(self,dev_dataloader, best_mrr, patience):
self.model.eval()
with torch.no_grad():
recall1_verb =[]
recall10_verb= []
mrr_verb = []
recall1_arg =[]
recall10_arg = []
mrr_arg = []
iterator = tqdm(dev_dataloader)
for batch in iterator:
source = batch["source"].to(self.device)
target = batch["target"].to(self.device)
generate_batch = self.model.model.generate(source,max_length=175, num_beams=10,num_return_sequences=10,early_stopping=True)
for i in range(len(target)):
gold_elem = self.tokenizer.decode(target[i], skip_special_tokens=True)
predictions = generate_batch[i*10:i*10+10]
new_predictions = []
for j in range(len(predictions)):
new_predictions.append(self.tokenizer.decode(predictions[j], skip_special_tokens=True))
def test_verb(gold_elem, predictions):
found = False
pattern = r"{(.*?)}"
gold_verb = re.findall(pattern, gold_elem, flags=0)[0].strip()
for idx,pred in enumerate(predictions):
pattern = r"{(.*?)}"
try:
pred_verb = re.findall(pattern, pred, flags=0)[0].strip()
except:
if idx==0:
recall1_verb.append(0.)
continue
if idx == 0:
if pred_verb == gold_verb:
recall1_verb.append(1.)
recall10_verb.append(1.)
mrr_verb.append(1.)
found=True
break
else:
recall1_verb.append(0.)
else:
if pred_verb == gold_verb:
recall10_verb.append(1.)
mrr_verb.append(1./float(idx+1))
found=True
break
if found ==False:
recall10_verb.append(0.)
mrr_verb.append(0.)
return mrr_verb,recall1_verb,recall10_verb
def test_arg(gold_elem, predictions):
found = False
pattern = r"{{(.*?)}}"
gold_arg = re.findall(pattern, gold_elem, flags=0)[0].strip()
for idx,pred in enumerate(predictions):
pattern = r"{{(.*?)}}"
try:
pred_arg = re.findall(pattern, pred, flags=0)[0].strip()
except:
if idx==0:
recall1_arg.append(0.)
continue
if idx == 0:
if pred_arg == gold_arg:
recall1_arg.append(1.)
recall10_arg.append(1.)
mrr_arg.append(1.)
found=True
break
else:
recall1_arg.append(0.)
else:
if pred_arg == gold_arg:
recall10_arg.append(1.)
mrr_arg.append(1./float(idx+1))
found=True
break
if found ==False:
recall10_arg.append(0.)
mrr_arg.append(0.)
return mrr_arg,recall1_arg,recall10_arg
mrr_v, rec1v, rec10v = test_verb(gold_elem, new_predictions)
mrr_a,rec1a, rec10a = test_arg(gold_elem, new_predictions)
metrics = {}
metrics['mrr_verbs'] = np.average(mrr_v)
metrics['recall@1_verbs'] = np.average(rec1v)
metrics['recall@10_verbs'] = np.average(rec10v)
metrics['mrr_args'] = np.average(mrr_a)
metrics['recall@1_args'] = np.average(rec1a)
metrics['recall@10_args'] = np.average(rec10a)
if best_mrr < metrics['mrr_verbs']:
best_mrr = metrics['mrr_verbs']
patience = 0
torch.save(self.model.state_dict(),self.model_path + "/bart_model_"+self.type_model+model_directory(self.model_path)+"_len"+str(self.max_len)+"_downsizing"+str(self.downsizing)+"k_action"+str(self.k_action)+"k_object"+str(self.k_object)+"_SEED10_lr2e-5.pt")
with open(self.model_path + "/best_mrr_"+self.type_model+".txt", "w") as f:
f.write(str(best_mrr))
with open(self.model_path + "/patience_"+self.type_model+".txt", "w") as f:
f.write(str(patience))
print(f"\033[1m Performance improvement, model saved in {self.model_path} \033[0m \n")
else:
patience +=1
with open(self.model_path + "/patience_"+self.type_model+".txt", "w") as f:
f.write(str(patience))
return metrics, best_mrr, patience
def prediction_final(self,test_dataloader):
self.model.eval()
with torch.no_grad(), torch.cuda.amp.autocast(enabled=True):
recall1_verb =[]
recall10_verb= []
mrr_verb = []
recall1_arg =[]
recall10_arg = []
mrr_arg = []
n = 45 if self.max_len == 20 else 200
iterator = tqdm(test_dataloader)
for batch in iterator:
source = batch["source"].to(self.device)
target = batch["target"].to(self.device)
try:
generate_batch = self.model.model.generate(source, max_length=n, num_beams=100,num_return_sequences=100, early_stopping=True)
except:
continue
for i in range(len(target)):
verbs_pred = []
args_pred = []
gold_elem = self.tokenizer.decode(target[i], skip_special_tokens=True)
predictions = generate_batch[i*100:i*100+100]
new_predictions = []
for j in range(len(predictions)):
new_predictions.append(self.tokenizer.decode(predictions[j], skip_special_tokens=True))
def find_verbs_args(gold_elem,new_predictions, verbs_pred, args_pred):
for idx,pred in enumerate(new_predictions):
pattern = r"{(.*?)}"
try:
pred_verb = re.findall(pattern, pred, flags=0)[0].strip()
if pred_verb not in verbs_pred:
verbs_pred.append(pred_verb)
except:
continue
for idx,pred in enumerate(new_predictions):
pattern = r"{{(.*?)}}"
try:
pred_arg = re.findall(pattern, pred, flags=0)[0].strip()
if pred_arg not in args_pred:
args_pred.append(pred_arg)
except:
continue
return verbs_pred, args_pred
def test_verb(gold_elem, verbs_pred):
found = False
pattern = r"{(.*?)}"
gold_verb = re.findall(pattern, gold_elem, flags=0)[0].strip()
for idx,pred_verb in enumerate(verbs_pred):
if idx == 0:
if pred_verb == gold_verb:
recall1_verb.append(1.)
recall10_verb.append(1.)
mrr_verb.append(1.)
found=True
break
else:
recall1_verb.append(0.)
else:
if idx < 10:
if pred_verb == gold_verb:
recall10_verb.append(1.)
mrr_verb.append(1./float(idx+1))
found=True
break
else:
if pred_verb == gold_verb:
mrr_verb.append(1./float(idx+1))
recall10_verb.append(0.)
found=True
break
if found ==False:
recall10_verb.append(0.)
mrr_verb.append(0.)
return mrr_verb,recall1_verb,recall10_verb
def test_arg(gold_elem, args_pred):
found = False
pattern = r"{{(.*?)}}"
gold_arg = re.findall(pattern, gold_elem, flags=0)[0].strip()
for idx,pred_arg in enumerate(args_pred):
if idx == 0:
if pred_arg == gold_arg:
recall1_arg.append(1.)
recall10_arg.append(1.)
mrr_arg.append(1.)
found=True
break
else:
recall1_arg.append(0.)
else:
if idx < 10:
if pred_arg == gold_arg:
recall10_arg.append(1.)
mrr_arg.append(1./float(idx+1))
found=True
break
else:
if pred_arg == gold_arg:
mrr_arg.append(1./float(idx+1))
recall10_arg.append(0.)
found=True
break
if found ==False:
recall10_arg.append(0.)
mrr_arg.append(0.)
return mrr_arg,recall1_arg,recall10_arg
start = 200
num_ret=200
verbs_pred, args_pred = find_verbs_args(gold_elem, new_predictions, verbs_pred, args_pred)
while len(args_pred)<10:
new_predictions = []
try:
pred_new = self.model.model.generate(source[i].unsqueeze(0), max_length=n, num_beams=start,num_return_sequences=num_ret, early_stopping=True)
except:
break
for j in range(len(pred_new)):
new_predictions.append(self.tokenizer.decode(pred_new[j], skip_special_tokens=True))
start += 100
num_ret +=100
verbs_pred, args_pred = find_verbs_args(gold_elem, new_predictions, verbs_pred, args_pred)
mrr_v, rec1v, rec10v = test_verb(gold_elem, verbs_pred)
mrr_a,rec1a, rec10a = test_arg(gold_elem, args_pred)
return mrr_v, rec1v, rec10v, mrr_a,rec1a, rec10a